Nowadays, continuous integration (CI) is indispensable in the software development process. A central promise of adopting CI is that new features or bug fixes can be delivered more quickly. A recent repository mining study by Bernardo, da Costa & Kulesza (2018) found that only about half of the investigated open source projects actually deliver pull requests (PR) faster after adopting CI, with small effect sizes. However, there are some concerns regarding the methodology used by Bernardo et al., which may potentially limit the trustworthiness of this finding. Particularly, they do not explicitly control for normal changes in the pull request delivery time during a project's lifetime (independently of CI introduction). Hence, in our work, we conduct a conceptual replication of this study. In a first step, we replicate their study results using the same subjects and methodology. In a second step, we address the same core research question using an adapted methodology. We use a different statistical method (regression discontinuity design, RDD) that is more robust towards the confounding factor of projects potentially getting faster in delivering PRs over time naturally, and we introduce a control group of comparable projects that never applied CI. Finally, we also evaluate the generalizability of the original findings on a set of new open source projects sampled using the same methodology. We find that the results of the study by Bernardo et al. largely hold in our replication. Using RDD, we do not find robust evidence of projects getting faster at delivering PRs without CI, and we similarly do not see a speed-up in our control group that never introduced CI. Further, results obtained from a newly mined set of projects are comparable to the original findings. In conclusion, we consider the replication successful.
Keywords: Continuous integration; Mining software repositories; Pull-request based development; Replication.
©2019 Guo and Leitner.